jev-pi-skills

TypeSafe Jev decisions for pi: small, structured calls (~0.4 s, fractions of a cent) that replace a frontier model deliberating over a closed question. Pure pi. It talks to Jev's API directly, with no other CLI or agent host.

| | what | when | |---|---|---| | jev_search | tool | After a web or API search: which results to open, whether the evidence is already enough, and which of your candidate queries to run next. Flags injected results. | | jev_rerank | tool | Before reading retrieved passages (memory, notes, history, wiki, web): relevance per passage, "is it answerable", and prompt-injection flags. Up to ~480 passages, in parallel batches. | | jev_supervise | tool | Checking a delegated or background run: keep_waiting, answer_question, nudge, escalate or collect, from the goal and the output tail. | | jev_pick | tool | One choice between options you define, e.g. which model or frontier seat takes a hard task. | | skill select | automatic | Before each turn, Jev ranks your whole skill catalog (batches of 120, then verifies the finalists) and names the skills this turn needs in the system prompt, or none. | | compaction | automatic | Before pi summarizes, Jev marks each assistant/tool turn keep, summarize or drop, and confidently dropped turns are removed. pi's own summarizer then runs as usual. User turns are never dropped. |

Companions: jev-pi-model-router (Jev picks the model per turn) and jev-pi-orchestrator (Jev-supervised multi-agent conductor). All three are independent.

Install

pi install npm:jev-pi-skills

Jev key from the console, first match wins: TYPESAFE_API_KEY env, then TYPESAFE_API_KEY= in the nearest .env, then ~/.pi/agent/secrets/typesafe_api_key (chmod 600). Restart pi.

/jevs                          status
/jevs skills off               stop the per-turn skill hint
/jevs compaction off           let pi compact on its own

State: ~/.pi/agent/jev-pi-skills/state.json. Debug: JEVS_DEBUG=/path/to/log pi logs one line per automatic decision. It never logs prompts or keys.

Measured (2026-09-24, pi 0.85, jev-1.13.0, live in pi via Herdr)

• Skill select on a 93-skill catalog took 1.4 to 1.5 s per turn. "Explain TDD in two sentences, no tools" got no hint (no TDD skill is installed). "An interactive HTML page to review and annotate a plan" got lavish, and the agent named it.
• jev_rerank on 4 passages kept both correct ones (0.99) and flagged the injected one (0.98).
• jev_pick sent "lock-free queue in Rust" to frontier (0.96 vs 0.04).
• Compaction on an 18-message session dropped 3 turns ("Sure! Let me take a look"-style) and pi summarized 68k tokens into 3.4k characters.

Upstream measured that Jev's keep/drop marks beat recency, but that a handoff written from a digest built on them recalled less than one written from the plain tail. So this package does not build digests. It only removes confidently dropped turns and leaves summarizing to pi. Turn it off with /jevs compaction off if you see summaries lose detail.

Privacy

Jev reads redacted text only: keys, tokens, emails and card-like numbers are masked, and long text is clipped (turns to 700 characters, passages to 900, the prompt to 2,000). Tool-call arguments are never sent. Skill names and descriptions are sent for skill select.

Credits

Question wording and thresholds are adapted from hermes-jev-skills (Steve Darlow, MIT), where they were measured. This package re-implements them for pi and includes none of that project's code. See NOTICE. MIT licensed.

Develop

npm test